What has to be true for AI to do real work?
- Every business action exists as a callable operation, not only as a screen a human clicks.
- Every record has a stable identity so the AI can refer to the same job twice.
- Permissions are enforced server-side, so an AI acting for a user cannot exceed that user's rights.
- Every action is written to an audit log with the actor recorded.
- Destructive or costly actions require explicit confirmation.
Why can't this be added later?
Because the hard part is not the model — it is the surface the model is allowed to touch. Systems built for humans expose screens; systems built for AI expose operations. Adding operations to a mature product means rebuilding the parts of it that assumed a person was always in the loop.
That is why late additions usually stop at text generation. Text is the only thing that can be produced without touching the workflow engine.
How can a buyer test for it?
Ask three questions: can the AI complete a multi-step task, does the result appear in the normal record where staff work, and can you see an audit entry afterwards. If the answer to any of those is no, the AI is sitting beside the product rather than inside it.